Traditional authorship attribution treats the problem of identifying a text's author as a standard multi-class classification task. If you train a machine learning model on a closed set of five candidates, the model learns a decision boundary that partitions the entire feature space among those specific individuals.
The fundamental flaw in this approach occurs when the model encounters a text written by someone else entirely. Because a standard softmax layer must distribute its probability mass among the known classes, it is forced to confidently misattribute the unknown text to the "least dissimilar" known candidate. As Lucas et al. (2019) and Zhai et al. (2022) point out, this closed-set constraint fails in real-world scenarios like forensic linguistics, social media moderation, and criminal investigations, where the true author is frequently an unknown individual outside the training database (Solorio et al., 2011; Sanderson et al., 2006; Johansson, 2019).
To solve this, researchers have shifted the paradigm from closed-set classification to open-set authorship verification. Instead of asking "Which of these known suspects wrote this text?", open-set systems ask "Did this specific person write this text, or did it originate from an unknown external source?"
Verification as a One-Class Problem
The core intuition behind authorship verification is that it must be treated as a one-class classification problem. As Veenman (2013) and Hürlimann et al. (2015) argue, the target author represents the only well-defined class. The "negative" class—everyone else who has ever written a piece of text—is effectively infinite and impossible to sample comprehensively.
Closed-Set Attribution (Softmax Partitioning):
[ Author A ] | [ Author B ] | [ Author C ] <-- Space is fully divided.
Unknown text *must* fall into A, B, or C.
Open-Set Verification (Boundary Profiling):
.-----------------.
/ Target Author \ <-- Model learns a tight boundary.
( Style Profile )
\ /
'-----------------'
X <-- Unknown text falls outside the boundary and is rejected.
To build a model capable of rejecting all known candidates when necessary, systems must evaluate whether a text matches a specific stylistic profile in isolation (Stamatatos et al., 2014). One highly effective strategy for this is the impostor method. Instead of relying on a static negative class, the system dynamically samples a set of external "impostor" texts from the web or independent corpora. If the questioned document is significantly closer to the target author's profile than to the background noise of the impostors, the system verifies authorship; otherwise, it rejects the match.
Universal Representation and Metric Learning
While one-class verification works well for a single disputed document, it struggles to scale when you need to track millions of potential accounts across the internet—a task known as account linking (Khan et al., 2021). In these massive, open-world environments, you cannot train a custom verification model for every single user.
Instead, the solution is deep metric learning. The goal here is to train a universal encoder that maps any text into a continuous embedding space where distance corresponds directly to stylistic similarity.
The primary challenge in learning these embeddings is separating style from topic. If an author writes a quantum physics paper and a cooking blog, a naive model will cluster the physics paper with other physics papers written by different people. To circumvent this, Sawatphol et al. (2022) introduce topic-regularized representation learning, which forces the network to ignore content words and focus purely on stylistic markers like syntactic dependency chains, punctuation density, and character n-grams (Bartelds et al., 2019).
Once this embedding space is established, open-set attribution becomes a matter of similarity matching and ranking rather than classification. When a new text appears, its embedding is calculated and compared against existing records. Investigators can then determine if the sample belongs to a previously identified actor or represents a completely new entity, without ever needing to retrain the underlying model (Saxena et al., 2023).
Modern Frontiers: From Antiquity to Generative AI
This shift to open-set verification has proven vital across diverse domains, from resolving centuries-old historical mysteries to defending against synthetic text.
- Historical Text Analysis: When analyzing ancient or disputed texts where the true author might not be preserved in history, closed-set models fail by forcing an attribution to a small pool of known figures. By applying fine-tuned language models within a verification framework, researchers have successfully demonstrated that composite works like the Ars Rhetorica or texts attributed to Plutarch were actually written by anonymous, unrepresented authors (Schmidt et al., 2024; Yamshchikov et al., 2022).
- Large Language Model Detection: Traditional AI detectors treat synthetic text detection as a closed-set binary problem ("Human vs. GPT-4"). However, Bevendorff et al. (2025) show that these detectors collapse when they encounter new generative models or diverse human writing styles not seen during training. Reframing LLM detection as an open-set verification task—verifying whether a text genuinely fits the broad, open-ended profile of human writing—creates a much more robust and generalizable defense.
- Internet-Scale Attribution: When the candidate pool expands to include anyone who has ever published on the internet, direct verification becomes computationally impossible. To solve this, Tan et al. (2025) propose a two-stage framework: a hierarchical keyword search first narrows down the infinite open world to a manageable candidate pool, and a multi-perspective reasoning model then performs the final verification.
Key Foundational Work
Modality Specific Meta Features for Authorship Attribution in Web Forum Posts (2011) — contributed 2 concept elements to this direction.
Short Text Authorship Attribution via Sequence Kernels, Markov Chains and Author Unmasking: An Investigation (2006) — contributed 1 concept elements to this direction.